Towards Transparent and Explainable Attention Models